LSR needs to remember inserted instructions even in postinc mode, because

there could be multiple subexpressions within a single expansion which
require insert point adjustment. This fixes PR7306.

llvm-svn: 105510

3 files changed, 53 insertions, 2 deletions

diff --git a/llvm/include/llvm/Analysis/ScalarEvolutionExpander.h b/llvm/include/llvm/Analysis/ScalarEvolutionExpander.h
index baf6946b8cf..656219d3229 100644
--- a/llvm/include/llvm/Analysis/ScalarEvolutionExpander.h
+++ b/llvm/include/llvm/Analysis/ScalarEvolutionExpander.h
@@ -32,6 +32,7 @@ namespace llvm {
     std::map<std::pair<const SCEV *, Instruction *>, AssertingVH<Value> >
       InsertedExpressions;
     std::set<Value*> InsertedValues;
+    std::set<Value*> InsertedPostIncValues;
 
     /// PostIncLoops - Addrecs referring to any of the given loops are expanded
     /// in post-inc mode. For example, expanding {1,+,1}<L> in post-inc mode
@@ -102,6 +103,10 @@ namespace llvm {
     /// clearPostInc - Disable all post-inc expansion.
     void clearPostInc() {
       PostIncLoops.clear();
+
+      // When we change the post-inc loop set, cached expansions may no
+      // longer be valid.
+      InsertedPostIncValues.clear();
     }
 
     /// disableCanonicalMode - Disable the behavior of expanding expressions in
@@ -146,7 +151,7 @@ namespace llvm {
     /// inserted by the code rewriter.  If so, the client should not modify the
     /// instruction.
     bool isInsertedInstruction(Instruction *I) const {
-      return InsertedValues.count(I);
+      return InsertedValues.count(I) || InsertedPostIncValues.count(I);
     }
 
     Value *visitConstant(const SCEVConstant *S) {
diff --git a/llvm/lib/Analysis/ScalarEvolutionExpander.cpp b/llvm/lib/Analysis/ScalarEvolutionExpander.cpp
index 0012b84e1f0..4f49bf17fb4 100644
--- a/llvm/lib/Analysis/ScalarEvolutionExpander.cpp
+++ b/llvm/lib/Analysis/ScalarEvolutionExpander.cpp
@@ -1312,7 +1312,9 @@ Value *SCEVExpander::expand(const SCEV *S) {
 }
 
 void SCEVExpander::rememberInstruction(Value *I) {
-  if (PostIncLoops.empty())
+  if (!PostIncLoops.empty())
+    InsertedPostIncValues.insert(I);
+  else
     InsertedValues.insert(I);
 
   // If we just claimed an existing instruction and that instruction had
diff --git a/llvm/test/CodeGen/X86/lsr-delayed-fold.ll b/llvm/test/CodeGen/X86/lsr-delayed-fold.ll
index 8afbb0d7a36..8ed97e447fe 100644
--- a/llvm/test/CodeGen/X86/lsr-delayed-fold.ll
+++ b/llvm/test/CodeGen/X86/lsr-delayed-fold.ll
@@ -132,3 +132,47 @@ for.inc131:                                       ; preds = %for.body123, %for.b
 for.end134:                                       ; preds = %for.inc131
   ret void
 }
+
+; LSR needs to remember inserted instructions even in postinc mode, because
+; there could be multiple subexpressions within a single expansion which
+; require insert point adjustment.
+; PR7306
+
+define fastcc i32 @GetOptimum() nounwind {
+bb:
+  br label %bb1
+
+bb1:                                              ; preds = %bb1, %bb
+  %t = phi i32 [ 0, %bb ], [ %t2, %bb1 ]      ; <i32> [#uses=1]
+  %t2 = add i32 %t, undef                     ; <i32> [#uses=3]
+  br i1 undef, label %bb1, label %bb3
+
+bb3:                                              ; preds = %bb1
+  %t4 = add i32 undef, -1                       ; <i32> [#uses=1]
+  br label %bb5
+
+bb5:                                              ; preds = %bb16, %bb3
+  %t6 = phi i32 [ %t17, %bb16 ], [ 0, %bb3 ]  ; <i32> [#uses=3]
+  %t7 = add i32 undef, %t6                    ; <i32> [#uses=2]
+  %t8 = add i32 %t4, %t6                    ; <i32> [#uses=1]
+  br i1 undef, label %bb9, label %bb10
+
+bb9:                                              ; preds = %bb5
+  br label %bb10
+
+bb10:                                             ; preds = %bb9, %bb5
+  br i1 undef, label %bb11, label %bb16
+
+bb11:                                             ; preds = %bb10
+  %t12 = icmp ugt i32 %t7, %t2              ; <i1> [#uses=1]
+  %t13 = select i1 %t12, i32 %t2, i32 %t7 ; <i32> [#uses=1]
+  br label %bb14
+
+bb14:                                             ; preds = %bb11
+  store i32 %t13, i32* null
+  ret i32 %t8
+
+bb16:                                             ; preds = %bb10
+  %t17 = add i32 %t6, 1                       ; <i32> [#uses=1]
+  br label %bb5
+}